HomeRoadmaps › Agentforce Specialist

Active Salesforce certification · Spring '26

Salesforce Certified Agentforce Specialist AI-201 Roadmap

Move from governed prompts and Data 360 retrieval to bounded agents, repeatable testing, production observability, and justified MCP or A2A orchestration.

105 minutes60 scored MCQUp to 5 unscored72% listed passing scoreNo prerequisite listed
Release-sensitive exam notice: Salesforce's official exam guide states that AI-201 questions align to Spring '26. This roadmap was checked on August 21, 2026. Recheck the official exam guide and registration page because releases, features, fees, and policies can change. No live or recalled questions are used.

Exact 50-question study allocation

The bank converts the six requested domains into exact counts: Prompt Engineering 10; Data 360 Fundamentals 10; AI Agents 17; Testing, Deployment, and Maintenance 5; Governance and Observability 5; Multi-Agent Orchestration 3. AI Agents deserves the largest practice block, but prompt and retrieval design together account for another 40% of this set.

Prompt Engineering · 10
Template fit, types, grounding, access, lifecycle, prompt quality, Trust Layer, model controls.
Data 360 · 10
Data Library, chunking, indexing, retrievers, freshness, identity, relevance, minimization.
AI Agents · 17
Agent Script, hybrid reasoning, topics, actions, variables, channels, runtime context, Agent API.
Lifecycle · 5
Testing Center, evaluation design, sandbox deployment, template dependencies, maintenance.
Governance · 5
Monitoring, audit evidence, trust signals, optimization, controlled change.
Multi-Agent · 3
Architecture choice, MCP tool boundaries, A2A delegation.

Phase 1 — Prompt Builder and trustworthy generation

Learn when generation is appropriate and when deterministic automation is safer. Build field-generation and flex-style scenarios, define explicit output contracts, and test missing evidence. Trace the template lifecycle from draft and preview through activation and invocation.

  • Map user goals, record context, grounding sources, output structure, and review responsibility.
  • Test template access separately from record and field access.
  • Study secure retrieval, data masking, prompt defenses, zero-data-retention arrangements, toxicity signals, and audit controls.
  • Create a model allowlist and deny tests for unapproved models.
  • Explain why better wording cannot replace authorization or server-side validation.

Phase 2 — Data 360 grounding and retrieval

Build a mental model from source ingestion to answer evidence. A useful Data Library needs curated content, coherent chunks, a current index, a correctly scoped retriever, permission-aware access, and an explicit no-result path.

  • Compare oversized, undersized, and coherent chunks against representative questions.
  • Measure source-to-index freshness and identify stale-content behavior.
  • Test synonyms, no-answer requests, conflicting documents, access denial, and identity-resolution mistakes.
  • Separate source quality, indexing, retrieval, prompt, and model failures.
  • Retrieve only fields and passages necessary for the task.

Phase 3 — Agent design and hybrid reasoning

Organize agents around coherent topics and bounded actions. Practice the next-generation authoring concepts named by the Spring '26 outline: Agent Script, Canvas and Script View, hybrid reasoning, filters, variables, and template expressions.

  • Use flexible conversation for clarification and deterministic gates for identity, eligibility, confirmation, and writes.
  • Prefer standard actions when they satisfy the requirement; make custom actions typed, least-privilege, idempotent, and auditable.
  • Test the actual runtime security context rather than relying on administrator preview.
  • Compare Employee and Service agents from audience and channel requirements.
  • Evaluate digital experience, email, voice, Slack, and Agent API integration concerns separately.

Phase 4 — Testing, deployment, and observability

Use Testing Center as a regression discipline. A useful suite specifies expected topic, evidence, action decision, output, side effect, and escalation. Include positive, negative, permission, ambiguous, stale-source, harmful-input, and tool-failure cases.

  • Inventory agent, topic, script, prompt, action, flow, permission, object, model, retriever, and channel dependencies.
  • Deploy to an inactive target, validate environment mappings, run tests under production-like identities, then canary and activate.
  • Monitor outcome, action success, escalation, error, latency, feedback, safety, and cost signals.
  • Keep audit access and retention bounded; do not create an uncontrolled conversation archive.
  • Practice rollback after an intentional action or dependency failure.

Phase 5 — Multi-agent architecture and final readiness

Begin with a single-agent baseline. Introduce specialist agents only when separation improves ownership, scale, or control enough to justify extra latency, cost, delegation, and failure complexity. MCP exposes tools and context; A2A supports inter-agent communication. Neither protocol creates trust automatically.

  • Define capability contracts, independent identities, minimized context, allowlisted tools, timeouts, structured errors, and trace correlation.
  • Inject unavailable specialists, malformed tool output, conflicting results, and revoked credentials.
  • Complete all three AI-201 projects.
  • Review all 40 AI-201 flashcards.
  • Run all 50 AI-201 practice questions under 105 minutes and explain every distractor.

Official-source study set

All AI-201 learning surfaces

Frequently asked questions

What is the AI-201 exam format?

The Spring '26 official guide lists 60 multiple-choice questions plus up to five unscored questions and 105 minutes. Unscored items are mixed into the exam and do not affect the result.

What passing score is listed?

The guide lists 72%. Verify the current official guide at scheduling time.

Is another Salesforce certification required?

No prerequisite is listed. Platform Administrator and Platform App Builder are described as recommended, not required.

Which release should I study?

The current guide states Spring '26. Use release-compatible Trailhead and Help content and recheck before the exam.

How much hands-on experience is recommended?

The audience description says successful candidates typically have one year of Salesforce platform and standard-object experience, including Data 360, Agentforce Builder, Prompt Builder, Testing Center, and sandbox-to-production patterns.

Are the practice questions recalled from the exam?

No. They are original scenarios written from the public outline and official documentation.

Start the complete AI-201 path

Read the guide, retrieve the flashcards, answer the exact 50-question bank, and complete the three controlled labs.

Read the study guide · Start questions · Review cards · Build projects